Conventional transmission control protocol/internet protocol (TCP/IP) offload engines residing on network interface cards (NICs) or elsewhere in a system such as in system software stacks, may inefficiently handle out-of-order (OOO) transmission control protocol (TCP) segments. For example, some conventional offload engines may merely drop out-of-order TCP segments. Dropped TCP segments need to be retransmitted by the sender, thereby utilizing additional bandwidth and reducing effective throughput. On links with large bandwidth-delay products such as high-speed local area network (LAN) of the order of 1 Gbps or faster, a large number of segments may be in transit between the sender and the receiver when the out-of-order TCP segment is dropped. Accordingly, many of the segments in transit must be retransmitted, thereby creating a substantial delay and excessive consumption of additional, often expensive and scarce bandwidth. TCP may also cut back on bandwidth allowed for a connection as the retransmission may be interpreted as being the result of congestion. This may further cause congestion avoidance mechanism to commence operation. A similar or even worse situation may arise with, for example, metropolitan area networks (MANs) with high bandwidth and moderate latencies or with long-haul wide area networks (WANs) that may have moderate bit rates and typical delays of the order of about 100 ms. In these types of networks, for example, system performance and throughput may be drastically reduced by the retransmissions.
In some conventional systems, on the sender or transmitter side, TCPs generally begin transmission by injecting multiple TCP segments into the network corresponding to a maximum window size that may be indicated by a receiver. In networks in which traffic traverses multiple networking entities or devices having varying link speeds, some of the networking entities or devices may have to queue TCP segments in order to handle the traffic. For example, network devices such as routers especially interfacing faster links with slower links in the communication path between the transmitter side and the receiver side may have to queue TCP segments. In this regard, there may be instances when there is insufficient memory on the networking entities or devices for queuing the TCP segments resulting in dropped segments. Accordingly, the TCP segments will have to be retransmitted, thereby consuming additional bandwidth.
In certain systems, retransmission may trigger TCP slow start and congestion-avoidance procedures which may result in substantial decrease in available bandwidth of a communication link. TCP slow start is an algorithm that may be utilized to minimize the effects of lost packets that may result from insufficient memory on slower networking entities or devices. TCP slow start utilizes a congestion window that is initialized to one TCP segment at the time of link initiation. In operation, the number of TCP segment allowed to be transmitted before and acknowledgment is received is incremented by one (1) for every acknowledgement (ACK) received from the remote peer. The sending side may therefore transmit a minimum number of TCP segments as specified by the minimum of the congestion window and the window that may be advertised by the receiving side. This may provide a near exponential growth in the window side and at some point, maximum capacity may be reached and the networking entity or device may start dropping packets.
Congestion avoidance is an algorithm that may be utilized in conjunction with slow start to minimize the effects of lost packets. Congestion may occur when a device may receive more TCP segments at its input than it may be able to adequately process or more then it can send on the egress. Congestion may also occur when TCP segments transition from a faster transport infrastructure to a slower transport infrastructure. In this regard, the network device at the edge of the faster transport infrastructure and the slower transport infrastructure becomes a bottleneck. Congestion avoidance utilizes packet loss and duplicate acknowledgements (ACKs) to determine when congestion occurs. As a result, the sender rate may be cut by half every time congestion is experienced.
Although slow start and congestion avoidance have varying objectives and are independent of each other, TCP recovery from congestion may involve decreasing the transmission rate and executing slow start to gradually increase the transmission rate from a window size of one (1). In some cases, TCP on the remote peer generates numerous ACKs and the local peer's congestion avoidance may interpret this to mean that TCP segments are lost, resulting in retransmission. Accordingly, TCP recovery from congestion avoidance and/or TCP slow start can be a relatively slow process especially for high bandwidth and may in certain instances, also cause unwanted retransmissions.
Other conventional offload engines may store out-of-order TCP segments in dedicated buffers attached to the offload engines residing on the MC or a host memory until all the missing TCP segments have been received. The offload engine may then reorder and process the TCP segments. However, storing the TCP segments in dedicated buffers can be quite hardware intensive. For example, the size of the dedicated buffers scale with the product of the bandwidth of the connections times the delay on the connections, and with the number of connections. In addition, storing the out-of-order segments on dedicated buffers may consume precious processor bandwidth when the out-of-order segments have to be reordered and processed. In addition, the offload engine still needs to handle other segments arriving at wire speed. Therefore, the reordering and processing may have to occur at the expense of delaying the processing of currently received TCP segments or by over provisioning of processing power that is scarce and hard to acquire for high speed of networks.
Accordingly, the computational power of the offload engine needs to be very high or at least the system needs a very large buffer to compensate for any additional delays due to the delayed processing of the out-of-order segments. When host memory is used for temporary storage of out-of-order segments, additional system memory bandwidth may be consumed when the previously out-of-order segments are copied to respective buffers. This choice complicates the processing of the data as the offload engine needs to communicate the state variables to a software agent for processing. While the software processes the state variables, the offload engine can't process new frames received for that TCP flow and has to buffer them. When the software agent is done, it needs to move the state variables back to the offload engine. If on the other hand, the offload engine tries to process the data stored on the host memory instead of the software agent, it encounters longer latencies than when processing frames locally, making this option very low performance or almost impractical.
Another design approach to a TCP offload Engine may be a flow-through approach. In the flow-through approach, the engine processes every TCP segment upon reception with no buffering, except for speed matching. The advantages of such a design approach are evidently the lack of external data buffering which scales with bandwidth delay product and with the number of connections. It adds cost, real estate and power to the solution as well additional pins on the offload engine ASIC to connect to the memory over a high speed bus. It also saves the additional complexity for reordering the out-of-order segments and processing them while additional traffic is received.
However, one challenge generally faced by TCP implementers wishing to design a flow-through NIC, is that TCP segments may arrive out-of-order with respect to the order in which they were transmitted. This may prevent or otherwise hinder the immediate processing of TCP control data and prevent the placing of the data in a host buffer. Accordingly, an implementer may be faced with the option of dropping out-of-order TCP segments or storing the TCP segments locally on the NIC until all the missing segments have been received. Once all the TCP segments have been received, they may be reordered and processed accordingly. In instances where the TCP segments are dropped or otherwise discarded, the sending side may have to re-transmit all the dropped TCP segments and in some instances, may result in about a fifty percent (50%) or greater decrease in throughput or bandwidth utilization, as described above.
Further limitations and disadvantages of conventional and traditional approaches will become apparent to one of skill in the art, through comparison of such systems with some aspects of the present invention as set forth in the remainder of the present application with reference to the drawings.